Skip to main content
Back to timeline
Hugging Face BlogSource publication:

Rumble-2-Flash opens at 0.5B parameters: 6.96% average error across nine languages, matching a 30B model on phone calls

Synopsis

Reson8 has open-sourced Rumble-2-Flash, a 0.5B-parameter speech recognition model under Apache 2.0 that covers 40+ languages with automatic language detection. On the Open ASR Leaderboard multilingual track it averages a 6.96% error rate across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English, below Meta's 15x larger omniASR-LLM 7B and ahead of Whisper large-v3; on a 302-hour call-centre benchmark with 14 English accents it reaches 7.86% WER, on par with NVIDIA's 30B Nemotron 3.5 Transcribe at 7.78%.

AI-generated editorial illustration: Rumble-2-Flash: an open 0.5B speech model for 40+ languages

Interpretation

A 0.5B-parameter speech recognition model now ships public weights under Apache 2.0, covering 40+ languages with automatic language detection and offering output controls for digits versus words, fillers, verbatim repeats and forced language. Multilingual recognition of this kind has typically been carried by much larger models; Reson8 says Rumble-2-Flash competes with models 3 to 60 times its size. On the Open ASR Leaderboard multilingual track the model averages a 6.96% error rate across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English, below Meta's omniASR-LLM 7B and ahead of Whisper large-v3; on the English track it averages 6.6% WER.

The model is obtained by distilling the larger production model Rumble-2 into a 0.5B student, in order to find how small a model could get before accuracy dropped off. On-policy distillation normally requires teacher and student to share a tokenizer, but choosing the Qwen3-ASR architecture meant taking its vocabulary too, and the two vocabularies differ. GOLD merges tokens where the two tokenizers split text differently so both models are compared over the same text: tokens in both vocabularies are compared directly and the rest by sorted probabilities; the student transcribes the audio itself and the teacher gives its own next-token probability at every position for the student to match token by token.

Throughput depends on how full the batches are, reaching over 7,000 RTFx on the larger datasets where batches stay full, about an hour of audio in half a second. That speed figure belongs to the larger datasets in HF's Open ASR Leaderboard where batches stay full, not to every workload. On AppTek's 302-hour call-centre benchmark, which covers 14 English accents and is scored with AppTek's own script, Rumble-2-Flash reaches 7.86% WER, on par with NVIDIA's Nemotron 3.5 Transcribe (30B) at 7.78%, ahead of the same-family Qwen3-ASR 0.6B (8.7%) and close to Qwen3-ASR 1.7B (7.4%).

Perspective

The demonstrated scope is the public leaderboard's multilingual and English tracks plus AppTek's 302-hour call-centre benchmark with 14 English accents; the weights are released under Apache 2.0 on the Hub and the model is also available through Reson8's API. The authors say they want user feedback to plan a next version.

The throughput figure depends on batch fullness, and the authors give over 7,000 RTFx only when batches stay full; the phone benchmark uses AppTek's own scoring script, and in that comparison Qwen3-ASR 1.7B still does better at 7.4%.